Search

The ability to log tutorial interactions in comprehensive, longitudinal, fine-grained detail offers great potential for educational data mining--but what data is logged, and how, can facilitate or impede the realization of that potential. We propose guidelines gleaned over 15 years of logging, exploring, and analyzing millions of events from Project LISTEN's Reading Tutor and its predecessors. (Contains 1 figure.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

Data mining methods are successful in educational environments to discover new knowledge or learner skills or features. Unfortunately, they have not been used in depth with collaboration. We have developed a scalable data mining method, whose objective is to infer information on the collaboration during the collaboration process in a domain-independent way and to improve collaboration process management and learning in an open collaborative educational web environment. Thus, we used statistical indicators of learner's interactions in forums as the data source and a clustering algorithm to classify the data according to learner's collaboration. We showed the information on learner's collaboration to the tutor and learners to help them with collaboration process management. The experimental results support this method. (Contains 1 figure and 2 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

This paper presents an agent-based educational environment to teach argument analysis (ALES). The idea is based on the Argumentation Interchange Format Ontology (AIF)using "Walton Theory". ALES uses different mining techniques to manage a highly structured arguments repertoire. This repertoire was designed, developed and implemented by us. Our aim is to extend our previous framework proposed in order to i) provide a learning environment that guides student during argument learning, ii) aid in improving the student's argument skills, iii) refine students' ability to debate and negotiate using critical thinking. The paper focuses on the environment development specifying the status of each of the constituent modules. (Contains 12 figures.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

Representing domain knowledge is important for constructing educational software, and automated approaches have been proposed to construct and refine such models. In this paper, instead of applying automated and computationally intensive approaches, we simply start with existing hand-constructed transfer models at various levels of granularity and use them as a lens to examine student learning. Specifically, we are interested in seeing whether we can evaluate schools by examining the grain-size at which its students are best represented. Also, we are curious about whether different types of students are best represented by different transfer models. We found that better schools and stronger students are best represented by models with a fewer number of skills. Weaker students and schools are best represented, for our data, by models that allow no transfer of knowledge in between skills. Perhaps surprisingly, to accurately predict the level at which a student represents knowledge it is sufficient to know his standardized test score rather than indicators of socio economic status or his school. (Contains 1 figure, 3 tables, and 2 footnotes.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

Student modeling is a widely used approach to make inference about a student's attributes like knowledge, learning, etc. If we wish to use these models to analyze and better understand student learning there are two problems. First, a model's ability to predict student performance is at best weakly related to the accuracy of any one of its parameters. Second, a commonly used student modeling technique, knowledge tracing, suffers from having multiple sets of parameters providing equally good model fits. Furthermore, common methods for estimating parameters, including conjugate gradient descent and expectation maximization, suffer from finding local maxima that are heavily dependent on their starting values. We propose a technique that estimates Dirichlet priors directly from the data, and show that using those priors produces model parameters that provide a more plausible picture of student knowledge. Although plausibility is difficult to quantify, we employed external measures to show the parameter estimates were indeed improved, even if our model did not predict student behavior any more accurately. (Contains 3 figures and 5 tables.) [Additional funding was provided by the Fulbright Program. For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

Most of the emphasis on mining online assessment logs has been to identify content-specific errors. However, the pattern of general "consistency" is domain independent, strongly related to performance, and can itself be a target of educational data mining. We demonstrate that simple consistency indicators are related to student outcomes, and suggest how consistency might be used in an online assessment framework to provide scaffolding to help students in need. (Contains 2 figures and 2 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

In this paper, we present a study of a large corpus of student logic exercises in which we explore the relationship between two distinct measures of difficulty: the proportion of students whose initial attempt at a given natural language to first-order logic translation is incorrect, and the average number of attempts that are required in order to resolve the error once it has been made. We demonstrate that these measures are broadly correlated, but that certain circumstances can make a hard problem easy to fix, or an easy problem hard to fix. The analysis also reveals some unexpected results in terms of what students find difficult. This has consequences for the delivery of feedback in the Grade Grinder, our automated logic assessment tool; in particular, it suggests we should provide different kinds of assistance depending upon the specific "difficulty profile" of the exercise. (Contains 4 figures and 3 footnotes.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

This paper proposes a novel method for recommending books to pupils based on a framework called Edu-mining. One of the properties of the proposed method is that it uses only loan histories (pupil ID, book ID, date of loan) whereas the conventional methods require additional information such as taste information from a great number of users which is costly to obtain. To achieve this, the proposed method solves the book recommendation problem as a problem of loan date prediction, relying solely on loan histories. Experiments show that the proposed method achieves an accuracy of 60% and outperforms the method (weighted slope open collaborative filtering) used for comparison. In addition to the performance, the proposed method has the following two advantages: (i) it is inexpensive compared to the conventional methods and (ii) reading level is adjustable. (Contains 2 figures, 4 tables, and 1 footnote.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

This paper presents several methods to automatically detecting students' mental models in MetaTutor, an intelligent tutoring system that teaches students self-regulatory processes during learning of complex science topics. In particular, we focus on detecting students' mental models based on student-generated paragraphs during prior knowledge activation, a self-regulatory process. We describe two major categories of methods and combine each method with various machine learning algorithms. A detailed comparison among the methods and across all algorithms is also provided. The evaluation of the proposed methods is performed by comparing the prediction of the methods with human judgments on a set of 309 prior knowledge activation paragraphs collected from previous experiments with MetaTutor on college students. According to our experiments, a content-based method with word-weighting and Bayes Nets algorithm is the most accurate. (Contains 1 figure and 2 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

In Cognitive Tutors, student skill is represented by estimates of student knowledge on various knowledge components. The estimate for each knowledge component is based on a four-parameter model developed by Corbett and Anderson [Nb]. In this paper, we investigate the nature of the parameter space defined by these four parameters by modeling data from over 8000 students in four Cognitive Tutor courses. We conclude that we can drastically reduce the parameter space used to model students without compromising the behavior of the system. Reduction of the parameter space provides great efficiency gains and also assists us in interpreting specific learning and performance parameters. (Contains 4 figures.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

The ability to predict a student's performance could be useful in a great number of different ways associated with university-level learning. In this paper, a grammar guided genetic programming algorithm, G3P-MI, has been applied to predict if the student will fail or pass a certain course and identifies activities to promote learning in a positive or negative way from the perspective of Multiple Instance Learning (MIL). Computational experiments compare our proposal with the most popular techniques of MIL. Results show that G3P-MI achieves better performance with more accurate models and a better trade-off between such contradictory metrics as sensitivity and specificity. Moreover, it adds comprehensibility to the knowledge discovered and finds interesting relationships that correlate certain tasks and the time devoted to solving exercises with the final marks obtained in the course. (Contains 4 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

Specifying the criteria of a rubric to assess an activity, establishing the different quality levels of proficiency of development and defining weights for every criterion is not as easy as one a priori might think. Besides, the complexity of these tasks increases when they involve more than one lecturer. Reaching an agreement about the criteria and the levels of proficiency might be easier taking into account the abilities students must achieve according to the purpose of the subject. However, the disagreement about the weights of every criterion in an assessment rubric might easily appear. This paper focuses on the automatic weight adjustment for the criteria of a rubric. This fitting can be considered as a global perception that the whole group of lecturers have about the accuracy of solving an activity. Firstly, each lecturer makes a proposal of weights and then, from a set of pairs of students he/she globally expresses who of each pair has solved better the activity for which the rubric was designed. Secondly, an approach based on the pairwise learning is proposed in this work to obtain adequate weights for the criteria of a rubric. The system commits fewer errors than the lecturers and make them improve and reconsider some aspects of the rubric. (Contains 3 figures and 5 tables.) [Funding was provided by the Ministerio de Ciencia e Innovacion (MICINN). For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

E-Learning systems offer students innovative and attractive ways of learning through augmentation or substitution of traditional lectures and exercises with online learning material. Such material can be accessed at any time from anywhere using different devices, and can be personalized according to the individual student's needs, goals and knowledge. However, authoring and evaluation of this material remains a complex a task. While many researchers focus on the authoring support, not much has been done to facilitate the evaluation of e-Learning applications, which requires processing of the vast quantity of data generated by students. We address this problem by proposing an approach for detecting potential symptoms of low performance in e-Learning courses. It supports two main steps: generating the production rules of C4.5 algorithm and filtering the most representative rules, which could indicate low performance of students. In addition, the approach has been evaluated on the log files of student activity with two versions of a Web-based quiz system. (Contains 4 tables.) [Funding for this paper was provided by the Spanish Ministry of Science and Education. For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

In this study, we are interested to see the impact of self-discipline on students' knowledge and learning. Self-discipline can influence both learning rate as well as knowledge accumulation over time. We used a Knowledge Tracing (KT) model to make inferences about students' knowledge and learning. Based on a widely used questionnaire, we measured students' level of self-discipline. When we analyzed the relation of students' self-discipline with their knowledge attributes, we found that high self-discipline students had significantly higher initial knowledge, but there is no consistent relationship of learning while using the tutor. Moreover, higher self-discipline students seemed more careful with respect to making careless mistakes. (Contains 5 figures and 5 tables.) [Additional funding was provided by the Fulbright Program. For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

Students in introductory programming classes often articulate their questions and information needs incompletely. Consequently, the automatic classification of student questions to provide automated tutorial responses is a challenging problem. This paper analyzes 411 questions from an introductory Java programming course by reducing the natural language of the questions to a vector space, and then utilizing cosine similarity to identify similar previous questions. We report classification accuracies between 23% and 55%, obtaining substantial improvements by exploiting domain knowledge (compiler error messages) and educational context (assignment name). Our mean reciprocal rank scores are comparable to and arguably better than most scores reported in a major information retrieval competition, even though our dataset consists of questions asked by students that are difficult to classify. Our results are especially timely and relevant for online courses where students are completing the same set of assignments asynchronously and access to staff is limited. (Contains 1 figure and 2 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

Identification of significant differences in sets of data is a common task of data mining. This paper describes a novel visualization technique that allows the user to interactively explore and analyze differences in mean values of analyzed attributes. Statistical tests of hypotheses are used to identify the significant differences and the results are then presented using Hasse diagrams. The presented technique has been tested on real data coming from pedagogical tests focused on evaluation of mathematical skills of secondary school students in Czech Republic. The results show that the proposed tool provides comprehensible representation of data. (Contains 4 figures and 4 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

This paper describes a collaborative educational data mining tool based on association rule mining for the continuous improvement of e-learning courses allowing teachers with similar course's profile sharing and scoring the discovered information. This mining tool is oriented to be used by instructors non experts in data mining such that, its internal operation is transparent to the user and the instructor can be focused in to the analysis of the results and make decisions about how to improve e-learning courses. (Contains 5 figures.) [Funding was provided by the Spanish Department of Research. For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

One of the main problems faced by university students is to take the right decision in relation to their academic itinerary based on available information (for example courses, schedules, sections, classrooms and professors). In this context, this work proposes the use of a recommendation system based on data mining techniques to help students to take decisions on their academic itineraries. More specifically, it provides support for the student to better choose how many and which courses to enrol on, having as basis the experience of previous students with similar academic achievements. For this purpose, we have analyzed real data corresponding to seven years of student enrolment at the School of system Engineering at Universidad de Lima. Based on this analysis, a recommendation system was developed. (Contains 1 figure and 1 table.) [Funding was provided by Spanish Ministry of Science and Education. For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

A basic question of instructional interventions is how effective it is in promoting student learning. This paper presents a study to determine the relative efficacy of different instructional strategies by applying an educational data mining technique, learning decomposition. We use logistic regression to determine how much learning is caused by different methods of teaching the same skill, relative to each other. We compare our results with a previous study, which used classical analysis techniques and reported no main effect. Our results show that there is a marginal difference, suggesting giving students scaffolding questions is less effective at promoting student learning than providing them delayed feedback. Our study utilizes learning decomposition, an easier and quicker approach of evaluating the quality of ITS interventions than experimental studies. We also demonstrate the usage of computer-intensive approach, bootstrapping, for hypothesis testing in educational data mining area. (Contains 4 tables and 1 footnote.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

The monitoring and support of university freshmen is considered very important at many educational institutions. In this paper we describe the results of the educational data mining case study aimed at predicting the Electrical Engineering (EE) students drop out after the first semester of their studies or even before they enter the study program as well as identifying success-factors specific to the EE program. Our experimental results show that rather simple and intuitive classifiers (decision trees) give a useful result with accuracies between 75 and 80%. Besides, we demonstrate the usefulness of cost-sensitive learning and thorough analysis of misclassifications, and show a few ways of further prediction improvement without having to collect additional data about the students. Attributes in the pre-university dataset are appended. (Contains 4 tables and 1 footnote.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

To make learning process more effective, the educational systems deliver content adapted to specific user needs. Adequate personalization requires the domain of learning to be described explicitly in a particular detail, involving relationships between knowledge elements referred to as concepts. Manual creation of necessary annotations is in the case of larger courses a demanding task. In this paper we tackle a concept relationship discovery problem that is a step in adaptive e-course authoring process. We propose a method of automatic concept relationship discovery for an adaptive e-course. We present two approaches based on domain model graph analysis. We evaluate our method in the domain of programming. (Contains 2 figures and 1 table.) [Funding was provided by the Cultural and Educational Grant Agency of the Slovak Republic. For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

We seek to simplify the creation of intelligent tutors by using student data acquired from standard computer aided instruction (CAI) in conjunction with educational data mining methods to automatically generate adaptive hints. In our previous work, we have automatically generated hints for logic tutoring by constructing a Markov Decision Process (MDP) that holds and rates historical student work for automatic selection of the best prior cases for hint generation. This method has promise for domain-independent use, but requires that correct solutions be assigned high positive values by the CAI or an expert. In this research we propose a novel method for assigning prior values to student work that depends only on frequency of occurrence for the component steps, and compare how these values impact automatic hint generation when compared to our MDP approach. Our results show that the utility metric outperforms a classic MDP solution in selecting hints in logic. We believe this method will be particularly useful for automatic hint generation for ill-defined domains. (Contains 2 figures and 4 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

This paper proposes a learning based method that can automatically determine how likely a student is to give a correct answer to a problem in an intelligent tutoring system. Only log files that record students' actions with the system are used to train the model, therefore the modeling process doesn't require expert knowledge for identifying domain specific skills that are needed to solve the problem or students' possible solution methods etc. The model utilizes a set of performance features, problem features, time and mouse movement features and is compared to i) a model that utilizes performance and problem features, ii) a model that uses performance, problem and time features. In order to address data sparseness problem, a robust Ridge Regression algorithm is designed to estimate model parameters. An extensive set of experiment results demonstrate the power of using multiple types of evidence as well as the robust Ridge Regression algorithm. (Contains 3 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

This paper presents a model which can automatically detect a variety of student speech acts as students collaborate within a computer supported collaborative learning environment. In addition, an analysis is presented which gives substantial insight as to how students' learning is associated with students' speech acts, knowledge that will significantly influence how this model is utilized by running learning software. Within Piagetian theory, the cognitive conflict of ideas between students is seen as beneficial for learning. Which sorts of interpersonal behaviors lead to most effective learning, however, is open to debate, with some researchers arguing that cooperation is most effective and others arguing that interpersonal conflict is a natural part of collaborative learning. We find that, in fact, interpersonal conflict is associated with positive learning, a finding that must be taken into account, in designing interventions that rely upon detectors of students' speech acts in CSCL environments. (Contains 1 figure and 3 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

The purpose of this study is to investigate the consistency of students' behavior regarding their pace of actions over sessions within an online course. Pace in a session is defined as the number of logged actions divided by session length (in minutes). Log files of 6,112 students were collected, and datasets were constructed for examining pace rank consistency in three main situations: day/night sessions, beginning/end (for both situations, sessions of the same learning mode were taken), and a comparison between sessions from different learning modes. For each dataset, students were ranked twice, according to their pace in the two sub-groups, and these ranks were correlated. Results obtained with this study's data suggest that pace is sometimes not consistent, hence might not be considered as a characterizing measure for the whole learning period. A discussion of this study and further research is provided. (Contains 1 figure and 5 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

Researchers who make tutoring systems would like to know which sequences of educational content lead to the most effective learning by their students. The majority of data collected in many ITS systems consist of answers to a group of questions of a given skill often presented in a random sequence. Following work that identifies which items produce the most learning we propose a Bayesian method using similar permutation analysis techniques to determine if item learning is context sensitive and if so which orderings of questions produce the most learning. We confine our analysis to random sequences with three questions. The method identifies question ordering rules such as, question A should go before B, which are statistically reliably beneficial to learning. Real tutor data from five random sequence problem sets were analyzed. Statistically reliable orderings of questions were found in two of the five real data problem sets. A simulation consisting of 140 experiments was run to validate the method's accuracy and test its reliability. The method succeeded in finding 43% of the underlying item order effects with a 6% false positive rate using a p value threshold of less than = 0.05. Using this method, ITS researchers can gain valuable knowledge about their problem sets and feasibly let the ITS automatically identify item order effects and optimize student learning by restricting assigned sequences to those prescribed as most beneficial to learning. (Contains 4 figures, 5 tables, and 3 footnotes.) [Additional funding for this paper was provided by the U.S. Department of Education's Graduate Assistance in Areas of National Need (GAANN). For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

The purpose of this study is to empirically reveal strategies of students' organization of learning-related digital materials within an online personal information archive. Research population included 518 students who utilized the personal Web space allocated to them on the university servers for archiving information items, and data describing their directory hierarchies. Several variables for measuring folders size and depth were defined, and four of them were chosen as best representing different aspects of the user's archive structure. Then, as a result of cluster analysis of the students, four organization strategies emerged, refining the classical piling/filing classification: piling, one-folder filing, small-folders filing, and big-folder filing. Also, associations were found between the organization strategies and archive size, students' studies degree. A discussion of this study and further research is provided. (Contains 4 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

It has been recognized that in order to drive Intelligent Tutoring Systems (ITSs) into mainstream use by the teaching community, it is essential to support teachers through the entire ITS process: Design, Development, Deployment, Reflection and Adaptation. Although research has been done on supporting teachers through design to deployment of ITSs, there is surprisingly little discussion about support for teachers' Reflection--the ability to draw conclusions from ITS usage, and Adaptation--adapting the content to better meet the needs of students. We describe our work on developing analysis tools and methodologies that support reflection and adaptation by teachers. The work was done in the context of helping teachers understand student's behavior in Adaptive Tutorials by post-analysis of the system's data-logs. We used a hybrid solution--part of the data-mining effort is teacher driven and part is automated. We tested our approach by comparing the results of expert analysis of two Adaptive Tutorials with and without an automated Refinement Suggestion Tool, and found it to be a useful teacher's aid. By using this tool, teachers act as "action researchers", confirming or disproving their hypotheses about the best way to use ITS technology. (Contains 3 figures.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

Traditional data mining techniques have been extensively applied to find interesting patterns, build descriptive and predictive models from large volumes of data accumulated through the use of different information systems. The results of data mining can be used for getting a better understanding of the underlying educational processes, for generating recommendations and advice to students, for improving management of learning objects, etc. However, most of the traditional data mining techniques focus on data dependencies or simple patterns and do not provide a visual representation of the complete educational (assessment) process ready to be analyzed. To allow for these types of analysis (in which the process plays the central role), a new line of data-mining research, called "process mining", has been initiated. Process mining focuses on the development of a set of intelligent tools and techniques aimed at extracting process-related knowledge from event logs recorded by an information system. In this paper we demonstrate the applicability of process mining, and the ProM framework in particular, to educational data mining context. We analyze assessment data from recently organized online multiple choice tests and demonstrate the use of process discovery, conformance checking and performance analysis techniques. (Contains 6 figures and 4 footnotes.) [Funding was provided by the Dutch Science Foundation (NWO). For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

In educational research, a fundamental goal is identifying which skills students have mastered, which skills they have not, and which skills they are in the process of mastering. As the number of examinees, items, and skills increases, the estimation of even simple cognitive diagnosis models becomes difficult. We adopt a faster, simpler approach: cluster a "capability matrix" estimating each student's individual skill knowledge to generate skill set profile clusters of students. We complement this approach with the introduction of an automatic subspace clustering method that first identifies skills on which students are well-separated prior to clustering smaller subspaces. This method also allows teachers to dictate the size and separation of the clusters, if need be, for practical reasons. We demonstrate the feasibility and scalability of our method on several simulated datasets and illustrate the difficulties inherent in real data using a subset of online mathematics tutor data. (Contains 3 figures and 2 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

Recent research has suggested that differences between intelligent tutor lessons predict a large amount of the variance in the prevalence of gaming the system. Within this paper, we investigate whether such differences also predict how much students choose to go off-task, and if so, which differences predict how much off-task behavior will occur. We utilize an enumeration of the differences between intelligent tutor lessons, the Cognitive Tutor Lesson Variation Space 1.1 (CTLVS1.1), to identify 79 differences between tutor lessons, within 20 lessons from an intelligent tutoring system for Algebra. We utilize a machine-learned detector of off-task behavior to predict 58 students' off-task behavior within that tutor, in each lesson. Surprisingly, the best model predicting off-task behavior from lesson features contains only one feature: lessons that involve equation-solving. We discuss possible explanations for this finding, and further studies that could shed light on this relationship. (Contains 1 figure and 2 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]

A fundamental goal of educational research is identifying students' current stage of skill mastery (complete/partial/none). In recent years a number of cognitive diagnosis models have become a popular means of estimating student skill knowledge. However, these models become difficult to estimate as the number of students, items, and skills grows. There exist alternatives such as sum-scores and the capability matrix. While initial theoretical work on sum-scores has been done, the behavior of sum-scores and the capability matrix is not well understood with respect to each other or to estimates from cognitive diagnosis models. In this paper we compare the performance of the three estimates of student skill knowledge under a variety of clustering methods using simulated data with varying levels of missing values. (Contains 6 tables.) [For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (2nd, Cordoba, Spain, July 1-3, 2009)," see ED539041.]